The speech database from which all context features were extracted and on which all training and testing took place is from the Boston University radio news corpus, speaker f2b [4]. The corpus contains approximately 45 minutes of female American English news-reader speech. The data was divided such that one quarter of the utterances were held out as a test set. The database is labelled with segment, syllable, and word boundaries, including lexical stress location. In addition, the corpus is labelled with ToBI intonation labels. The database was labelled with Tilt intonation labels by automatically mapping the ToBI tone labels to Tilt event labels [3]. The Tilt parameters from these labels were used to train the decision trees for the experiment.